One-Shot Facial Recognition using Siamese Neural Networks on LFW Dataset

This project implements a one-shot facial recognition system using Siamese Neural Networks, trained and evaluated on the LFW-a (Labeled Faces in the Wild - aligned) dataset. The network is inspired by Koch et al.'s 2015 paper: "Siamese Neural Networks for One-shot Image Recognition", and is designed to verify whether two facial images depict the same person—even when only one example per identity is available.

---

📂 Project Overview

• One-shot Learning: Works with extremely limited training samples per identity.
• Siamese Network Architecture: Learns similarity between image pairs instead of categorical classification.
• Hyperparameter Tuning: Utilizes Optuna for automated, large-scale search.
• Data Augmentation: Improves model robustness using controlled transforms.
• Real-world Evaluation: Tested on LFW, a challenging facial dataset with high variation.

---

📊 Dataset: LFW-a

• Source: Aligned version of the LFW dataset.
• Samples: 13,233 facial images.
• Pairs: CSV-based file structure defining image pairs labeled as same (1) or different (0).
• Train/Validation/Test Split: Stratified and balanced, with 20% held out for validation.

---

🏗️ Model Architecture

Component	Description
Input	Grayscale images (105x105, 160x160, or 200x200)
Conv Layers	4 convolutional blocks with ReLU + MaxPooling (kernels: 10, 7, 4, 4)
Fully Connected	4096-unit dense layer with optional Dropout followed by a scalar output
Output	Absolute difference between embeddings → sigmoid → similarity score

---

🛠️ Preprocessing & Augmentation

• Cropping: Removes 30px border to reduce noise.
• Resizing: Evaluated across 105×105, 160×160, and 200×200.
• Normalization: Pixel values scaled to [-1, 1].
• Augmentations (Train only):
  • RandomHorizontalFlip
  • RandomRotation(±10°)

---

⚙️ Hyperparameter Tuning (via Optuna)

Parameter	Range / Choices
Image Size	105x105, 160x160, 200x200
Learning Rate	Log-uniform [1e-6, 1e-3]
Weight Decay	Log-uniform [1e-6, 1e-2]
Dropout Rate	Uniform [0.0, 0.5]
Optimizer	Adam, AdamW, RMSprop, SGD
Batch Size	32
Loss Function	Binary Cross Entropy
Regularization	L1

---

🧪 Results

Resize	Train Acc	Val Acc	Test Acc	Train Loss	Val Loss	Epochs
105×105	64.66%	63.64%	65.80%	0.6194	0.6242	36
160×160	75.91%	67.73%	68.00%	0.5065	0.5681	25
200×200	82.10%	71.82%	68.50%	0.4185	0.5637	25

• ⏱️ Early stopping patience: 20 epochs
• 📉 Best performance obtained at 200×200 input size, with fastest convergence and highest generalization.

---

🔍 Comparison with Koch et al. (2015)

Feature	Koch et al.	Our Implementation
Dataset	Omniglot	LFW-a
Input Image Size	105×105 RGB	105×105 to 200×200 grayscale
Optimizer	SGD	Adam, AdamW, RMSprop, SGD
Pretrained Models	Yes (in later works)	No (trained from scratch)
Accuracy (Best)	~92% (Omniglot)	~68.5% (LFW-a)

---

🧠 Future Improvements

• ✅ Use pretrained backbones like FaceNet or VGGFace.
• 🔄 Add Batch Normalization between Conv layers.
• 🎛️ Explore stochastic data augmentation for variability.
• 🧪 Tune dropout and regularization more aggressively for high-res inputs.

---

📁 Files

• code.ipynb — Full implementation and training pipeline.
• report.pdf — Formal write-up with detailed experiments, results, and analysis.

---

👥 Authors

• Gil Ari Agmon
• Shir Rozenfeld

Project submitted as part of Deep Learning – Assignment 2, Ben-Gurion University of the Negev.

---

🛡️ License

This project is for academic use only.